Target recognition method and system

Abstract

The invention provides a target recognition method and system. The method comprises the steps of firstly obtaining a to-be-recognized image; inputting the to-be-recognized image into a convolutional neural network model to obtain a recognition result output by the neural network model, wherein the convolutional neural network model is obtained by the following steps: training based on an image training sample carrying an object label, screening different channels under any convolutional layer based on the L1 norm of the floating point type parameters of different channels under any convolutional layer in the convolutional neural network model after training is finished, and performing quantization processing on the screened convolutional neural network model. Since different channels underany convolutional layer in the applied convolutional neural network model are screened, the floating-point number value ranges of different channels in the convolutional layer obtained after screening are close, the network precision of the quantized convolutional neural network model can be improved, the accuracy of an identification result output by the convolutional neural network model is further improved, and the availability of the convolutional neural network model is improved.

Description

technical field [0001] The present application relates to the technical field of machine learning, in particular to a target recognition method and system. Background technique [0002] Currently, the Convolutional Neural Networks (CNN) model is widely used in image detection, object recognition and other fields. In the field of target recognition, in order to improve the processing speed of the image to be recognized by the convolutional neural network model, it is usually necessary to convert the floating-point convolutional neural network model into a fixed-point convolutional neural network model. the quantification process. [0003] In the prior art, when quantizing the convolutional neural network model, there are usually the following two implementation methods: one is to perform channel-by-channel quantization for each convolutional layer, and this quantization method makes the quantized convolutional neural network The network precision of the network model is hig...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, when quantizing the convolutional neural network model, there are usually the following two implementation methods: one is to perform channel-by-channel quantization for each convolutional layer, and this quantization method makes the quantized convolutional neural network The network precision of the network model is high, but the process is complicated and not friendly to the processor; the other is to perform overall quantization for each convolutional layer, that is, quantize tensor by tensor. However, due to the different value ranges of floating-point numbers in different channels, the network accuracy of the quantized convolutional neural network model is low

Images